Abstract
Brain tumors pose a critical threat to human health, and early detection is essential for improving patient outcomes. This study presents two key enhancements to the YOLOv11 architecture aimed at improving brain tumor detection from MRI images. First, we integrated a set of novel attention modules (Shuffle3D and Dual-channel attention) into the network to enhance its feature extraction capability. Second, we modified the loss function by combining the Complete Intersection over Union (CIoU) with a Hook function (HKCIoU). Experiments conducted on a public Kaggle dataset demonstrated that our improved model reduced parameters and computations by 2.7% and 7.8%, respectively, while achieving mAP50 and mAP50-95 improvements of 1.0% and 1.4%, respectively, over the baseline. Comparative analysis with existing models validated the robustness and accuracy of our approach.